DroneNet: Network Reconstruction through SparseConnectivity Probing using Distributed UAVs

Dahee Jeong, So-Yeon Park, and HyungJune Lee†Department of Computer Science and Engineering

Ewha Womans University, Seoul, South KoreaEmail: {dahee.jeong, ewsy25}@i.ewha.ac.kr, hyungjune.lee@ewha.ac.kr

Abstract—In this paper, we consider a network reconstruc-tion problem using UAVs where stationary ad-hoc networksare severely damaged in a post-disaster scenario. The mainobjective of this paper is to repair network by supplementingaerial wireless links into the stationary network to reconnectisolated ground networks each other with a limited number ofUAVs. We propose a distributed motion planning that guaranteescomplete coverage to probe network connectivity from the airover stationary networks, while reducing duplicate coverage withother UAVs. Given the collected local connectivity informationover region of interest, we deploy UAVs as relays into the locationsof network holes to repair network-wide data delivery mosteffectively by formulating the problem into a binary integerprogram. Simulation results show that our network traversingalgorithm outperforms a multi-agent exploration algorithm Antsin terms of complete coverage time, travel distance, and duplicatecoverage. Also, our deployment optimization enhances network-wide routing performance compared to a practical baselinecounterpart.

I. INTRODUCTION

In catastrophic diaster scenarios, maintaining a reliablecommunication network through fast network repair is im-portant for effectively sharing in-situ emergency informationbetween victims and first responders. There have been effortson utilizing autonomous unmanned terrestrial or aerial vehicleson a Region of Interest (ROI) [4]. These mobile vehiclescan be used as effective communication resources to quicklyreconnect isolated networks each other through the ad-hocdeployment.

An advantage of unmanned aerial vehicles (UAVs) com-pared to terrestrial vehicles is its physically less constrainedmovement for information gathering. The UAVs can retrievedata from the ground via the ground-to-air communication,relay them to other UAVs in the air-to-air communication, andsend back to the ground via the air-to-ground communication.They can gather network collapse status with connectivityprobing from the air, and also be deployed by themselves ascommunication relays if necessary.

We consider the major roles of UAVs as exploring net-work connection status over unknown ROI areas, and beingdeployed as ground-to-air and air-to-ground relays for au-tonomous network reconstruction. The challenges are 1) todesign a distributed motion planning algorithm for sparse yetefficient connectivity probing over the damaged network, and

†HyungJune Lee is the corresponding author.

2) to locate network holes where the deployment of UAVrelays helps to repair the damaged network.

There have been previous works to address the problem ofmulti-agent exploration in [6], [9], [11] mostly from roboticsresearch community. In [9], [11], researchers propose simpledistributed Ants algorithms simulating a colony exploration ofants while leaving pheromone traces during the environmenttraversing. Although [6], [7] present the Brick&Mortar algo-rithm and its variation that reduces duplicate coverage as op-posed to the Ants, they suffer from somewhat computationallyintensive loop closure procedures.

The problem of network hole detection and deployment hasbeen studied in [1]–[3], [5], [10], [12] from network researchcommunity. [1], [3], [10] explore sensor deployment algo-rithms by finding network holes in sensor networks from moretheoretical perspectives. Regarding the usage of aerial vehicles,aerial communication based on 802.11n performs poorly dueto aerial link vulnerability [2], while some antenna extensioncan enhance the quality of aerial links [12]. Some researchersutilize UAVs to re-establish network connectivity with aerialdeployment in [5]. However, network repair improvement withrespect to network probing density and the optimal UAVdeployment problem based on sparse connectivity informationhave not been investigated well.

In this paper, we aim to answer two key questions of 1)how to traverse a network efficiently with multiple UAVs ina distributed way without much duplicate coverage and 2)what the optimal UAV deployment algorithm based on tangibleconnectivity probing measurements should be to achieve apractical network recovery.

We propose a novel distributed motion planning algorithmbased on independent and computationally light decisionsamong several pre-determined zigzag patterns. These patternsextend the local coverage as the UAVs are flying forward,while reducing duplicate coverage with other UAVs. Exploringthe ROI area, the UAVs periodically probe network connec-tivity from the air toward stationary networks.

Once the network traversing procedure is completed, wefind the optimal UAV relay positions that can repair network-wide data delivery most effectively. We formulate the probleminto a binary integer program and obtain the optimal deploy-ment strategy.

The rest of this paper is organized as follows: After de-scribing system model in Sec. II, we present our network

PHASE 1. Network traversing by distributed UAVs

vertex

PHASE 2. Coverage hole detection

Deployment of Multi-UAV Relays for network recovery

Optimization

Fig. 1. Overall procedure of network traversing, coverage hole detection, anddeployment by exploiting UAVs for autonomous network recovery.

traversing algorithm in Sec. III. In Sec. IV, we propose a UAVdeployment algorithm based on binary integer optimization.After we evaluate our algorithms in a simulation environmentin Sec. V, we finally conclude this paper in Sec. VI.

II. SYSTEM MODEL

This work considers a network reconstruction problemusing UAVs where stationary ad-hoc networks are severelydamaged in a post-disaster scenario. Our goal is to repairnetwork coverage by supplementing aerial wireless links intothe stationary network to reconnect isolated ground networkseach other with a limited number of UAVs.

We assume that UAVs are equipped with the same wirelessradio as stationary nodes (e.g., 802.11 or 802.15.4). A UAVcan communicate with a part of stationary nodes on the groundor other UAVs in the air as long as they are within radio range.It is also assumed that UAVs are aware of Region of Interest(ROI) to explore and can keep track of their relative positionon ROI compared to its corresponding physical position.Any UAV control issues on moving from one location toanother due to external environmental factors such as weather,obstacles, and collisions with other UAVs are out-of-scope inthis paper. We consider UAVs to initially be fully charged andkeep operating without recharging during a complete missionof network reconstruction.

The problem of network construction using UAVs canbe divided into two sub-problems: 1) network connectivityprobing from the air based on distributed motion planningof UAVs for the complete ROI coverage, while reducingduplicate coverage (refer to Sec. III), and 2) optimal UAV relaydeployment for the most effective network recovery given alimited number of UAVs (refer to Sec. IV), as in Fig. 1.

III. NETWORK TRAVERSING

In this section, we propose our novel distributed motionplanning algorithm of multiple UAVs and describe the proce-dure of network connectivity traversing of the UAVs. MultipleUAVs explore the network over Region of Interest (ROI)according to their own independent navigation decision. For anefficient distributed exploration on the ROI region, we define afrontier map that consists of square grids with m×m vertexesas in Fig. 2(a). Each UAV initiates its navigation at its currently

𝑆1 𝑆2 𝑆8⋯ ⋯

𝑆57

𝑆9

𝑆64

⋯⋯

𝑣8 𝑣9 𝑣10

𝑣18 𝑣19 𝑣20

𝑣28 𝑣29 𝑣30

𝑣1 𝑣2

𝑣78 𝑣79 𝑣80

𝑣88 𝑣89 𝑣90

𝑣98 𝑣99 𝑣100

𝑣71 𝑣72 𝑣73

𝑣81 𝑣82 𝑣83

𝑣91 𝑣92 𝑣93

𝐿

𝑁𝐸𝑁𝑊

𝑊𝑁

𝑊𝑆 𝐸𝑆

𝐸𝑁

𝑆𝑊 𝑆𝐸

𝐿

(a) Logical grid coordinate consisting of vertexes,also showing eight pre-determined zigzag patternsfor motion planning.

𝑣71

𝑣81

𝑣91 𝑣92 𝑣93

𝑣51

𝑣61

𝑣94 𝑣95

𝑣71

𝑣81

𝑣91 𝑣92 𝑣93

𝑣51

𝑣61

𝑣94 𝑣95

(b) Case 1: The East vertex towardthe first quadrant with the longestlength up to ROI is taken. Then, thepattern including the North vertex ischosen.

𝑣71

𝑣81

𝑣91 𝑣92 𝑣93

𝑣51

𝑣61

𝑣94 𝑣95

𝑣71

𝑣81

𝑣91 𝑣92 𝑣93

𝑣51

𝑣61

𝑣94 𝑣95

(c) Case 2: Both North and East ver-texes toward the first quadrant withthe longest length up to ROI aretaken. Then, it finds a next longestpattern on another quadrant.

Fig. 2. Logical grid coordinate, zigzag movement trajectories, and futurevertex visit decision rules.

visiting vertex or a designated vertex, continues its movementdecision to the next vertex, and stops if it covers all of thevertexes on the ROI.

Each UAV initially generates a future-vertex-visit-trajectorywith the longest length toward a certain direction up to theboundary of ROI among eight pre-determined zigzag patterns(North-East, North-West, South-East, South-West, East-North,East-South, West-North, West-South) as shown in Fig. 2(a).Whenever a UAV visits a vertex at a time, it adds the visitedvertex ID to its vertex-visit-list. If two or more UAVs arewithin radio range, they share their own vertex-visit-list withothers and merge them into its original vertex-visit-list.

When a UAV decides its next visiting vertex based on thefuture-vertex-visit-trajectory, it checks whether the anticipatingvisiting vertex has already been taken by other UAVs bysearching it over the vertex-visit-list. In case that the antic-ipating vertex is already taken, the UAV lists up all availableneighboring vertexes to move among (North, East, South,West), except the direction with the taken vertex, and randomlychoose one direction for next move. In this way, a UAV isable to avoid duplicate exploration over the vertexes alreadyvisited by other UAVs in a distributed manner. It continues togenerate a future-vertex-visit-trajectory with the longest lengthtoward the boundary of ROI and execute its local visit decision

Algorithm 1 Distributed Multi-UAV Network Traversing1: Input: CurrentVertexID2: Output: NextVertexID

// Part I: Motion planning3: if (future-vertex-visit-trajectory == ∅) then4: Regenerate the future-vertex-visit-trajectory

with the longest length that starts from an unvisited vertex;5: NextVertexID = future-vertex-visit-trajectory’s first vertex ID;6: Move with one step to the next vertex;7: else8: if future-vertex-visit-trajectory’s next vertex is taken or null then9: future-vertex-visit-trajectory = ∅;

10: if (any unvisited neighboring vertex in North, East, South, West)then

11: NextVertexID = random-pick(unvisited neighboring vertexes);12: Invoke connectivity-probing();13: Move with one step to the next vertex;14: else15: if (there exists any unvisited vertex) then16: NextVertexID = the nearest vertex’s ID on the grid coordinate

from CurrentVertexID;17: Invoke connectivity-probing();18: Fly to the next vertex;19: else20: Terminate;21: end if22: end if23: else24: NextVertexID = future-vertex-visit-trajectory’s next vertex ID;25: Invoke connectivity-probing();26: Move with one step to the next vertex;27: end if28: end if

// Part II: Connectivity probing29: Function connectivity-probing()30: Broadcast hello packets;31: Receive response packets from neighboring stationary nodes;32: Calculate the average PRR for each responded stationary node;33: Update the PRR table for CurrentVertexID and StatonaryNodeId;34: if (any UAVs within radio range) then35: Exchange vertex-visit-list and PRR table, and update them36: end if37: EndFunction

afterwards as illustrated in Figs. 2(b) and 2(c).During each vertex visit, a UAV probes network connec-

tivity with neighboring stationary nodes near the vertex. TheUAV broadcasts hello packets with a periodic manner, and anyneighboring stationary node that have received a hello packetreplies back to the UAV with a response packet embeddingits own node ID. Based on the collected response packetsfrom connectable nodes for multiple hello packets, the UAVcalculates the average Packet Reception Rate (PRR) for eachresponded node ID at the vertex position. As each UAVtraverses over the network on the ROI, it continuously updatesits PRR table for the attributes of visited vertex ID andstationary node ID and also exchanges its PRR table togetherwith the vertex-visit-list if other UAVs are within radio range.

If a UAV checks that all of neighboring vertexes in thenorth, east, south, and west directions are taken, it comparesits vertex-visit-list with the entire vertex list on ROI, andselects an unvisited vertex with the shortest distance on thegrid coordinate for its next move. In this case, the UAV directlyflies to the selected vertex. If there remains no vertex to visit,it finishes the network traversing procedure.

Our network traversing algorithm guarantees the completecoverage of vertexes with distributed motion planning of mul-tiple UAVs and its successful termination without overlappingloops. The proofs are straightforward and omitted due to spaceconstraints.

IV. UAV RELAY DEPLOYMENT

In this section, we present a UAV relay deployment algo-rithm that finds the best grid positions of multiple UAVs for theoptimal network repair. Given the collected local connectivityinformation over Region of Interest (ROI), we find criticalnetwork holes that drastically undermine network-wide routingperformance. We want to deploy a limited number of availableUAVs as relays into the locations where local connection aswell as end-to-end routing can significantly be improved.

Once the UAVs complete the network traversing procedurein Sec. III, we obtain the connectivity table consisting ofPacket Reception Ratio (PRR) from stationary node ID k atvertex ID j, i.e.,

[PRR

]j,k where 1 ≤ j ≤ M(= m2)

and 1 ≤ k ≤ N . To find the network holes, we observe aset of vertexes that retain the weakest wireless links to theneighboring stationary nodes. Since the number of availableUAVs is limited, we prioritize the network holes and selectsome of them as UAV relay deployment positions. It shouldbe noted that UAVs should not be deployed into the networkholes completely isolated by any neighboring stationary nodesbecause the deployed relay still remains unconnected to anyof them.

To benefit the overall network from only few UAV relaysfor network reconstruction, we aim to minimize duplicate net-work coverage by prohibiting two or more UAVs from beingdeployed within communication range. Thus, we want eachUAV to contribute to repairing its nearby network connectivitywithout partial or complete duplicate coverage for the overallnetwork repair enhancement.

We formulate the problem of selecting grid positions ofmultiple UAVs for global network repair into a binary integerprogram. Our goal is to find a set of vertex regions thathave the weakest non-zero PRRs averaged over neighboringstationary nodes, while avoiding duplicate coverage with anyof other UAVs.

To formulate this setting, we first define a group of vertexeson a square sub-grid within the average radio range of awireless interface as Si = {vi1 , vi2 , vi3 , . . . , vin2} where vil(1 ≤ l ≤ n2) is a vertex element belonging to the setSi, and n2 is the total number of elements in set Si, andS1 ∪S2 ∪ · · · ∪SK = {v1, v2, . . . , vm2} as in Fig. 2(a). Giventhe P number of UAVs to deploy, the problem of selecting Pgrid positions of UAVs is to select the P number of sets withthe lowest average PRRs over their corresponding belongingvertexes among S1, S2, . . . , and SK , while any selected vertexsets should not share any vertex in common. Both Si and Sj

cannot be selected if Si ∩ Sj 6= ∅. For example, in Fig. 2(a),both S1 and S2 cannot be selected as deployment vertexes.This implies that we want to deploy a UAV into a group

location of vertexes of which most or all suffer from similarlypoor connection.

We introduce indicator functions Ji denoting the vertexgroup set Si should be selected, and Ii,j denoting whetherthe vertex group set Si and its belonging vertex vj should beselected.

Based on these notations, we define the objective function tominimize the summation of the average PRRs of the selectedvertexes in the selected vertex set as follows.

minimize∑

i,j∈Si

PRRj · Ii,j (1)

subject to∑i

Ii,j ≤ 1 ∀j (2)

Ji = Ii,i1 = Ii,i2 = Ii,i3 = · · · = Ii,in2 ∀i (3)∑i

Ji = P (4)

where PRRj is the average PRR over only the stationarynodes with non-zero PRRs at vertex vj . In case that vertex vjhas no connection at all, i.e., PRRj = 0, we force it to be 1so that isolated vertexes should never be selected.

Constraint (2) ensures that any selected vertex sets shouldnot share any vertex in common to avoid duplicate coverage.Constraint (3) enforces the condition that once a vertex groupset Si is selected, any belonging vertex vj ∈ Si should beselected. The last constraint (4) requires the total number ofselected vertex group sets to be the same number of UAVs.

By using MATLAB bintprog utility or AMPL/CPLEXsolver, we can obtain the optimal sets of the most vulnerablevertex groups under critical link outage. Since each UAV endsup with the entire connectivity table for all the vertexes at theend of network traversing procedure, it calculates them foritself. Once each UAV tracks down to these sets, it determinesone of sets according to the order of UAV ID, and fliesdirectly to the center position of the selected vertex groupfor its self-deployment. These positions are exactly where thedeployed UAVs can be used as crucial relay resources forstarting repairing the broken network.

V. EVALUATION

We validate our proposed scheme in a simulated damagednetwork of 218 stationary sensor nodes in a 500× 500 m2 areawhere the networks are sparsely connected each other withsome degree of isolation. In particular, we focus on the ROIarea in a 280 × 280 m2 to which 65 stationary nodes belongas shown in Fig. 3(a). To simulate the radio propagation, weuse a combined path-loss shadowing model with a path-lossexponent of 3, a reference loss of 46.67 dB, and an additivewhite Gaussian noise N(0, 52) in dB [8].

We evaluate our network traversing algorithm comparedto Ants algorithm [11] in terms of complete coverage time,travel distance, and duplicate coverage rate with respect tothe number of UAVs in Sec. V-A. Then, we validate how ourdeployment algorithm improves the overall damaged network

−50 0 50 100 150 200 250 300−50

0

50

100

150

200

250

300

X (m)

Y (

m)

(a) Before deployment.

−50 0 50 100 150 200 250 300−50

0

50

100

150

200

250

300

X (m)

Y (

m)

(b) After deployment of 4 UAVs (denoted as red triangles) withDroneNet.

Fig. 3. Network connectivity of sensor nodes over ROI (within the red squareboundary) in a simulated network (where wireless links are shown for PRR≥ 75%).

in terms of end-to-end routing cost and routing hole fractionby varying the number of UAVs and the probing density ofnetwork traversing in Sec. V-B.

In our experiments, the total number of vertexes in gridover ROI, M is 100 where m = 10, and the total number ofelements in vertex set Si is 9 where n = 3 corresponding tothe average radio range (as illustrated in Fig. 2(a)). 4 UAVs(i.e., P = 4) are deployed in the height of 9 m in the air.Each UAV calculates its PRR table based on 50 hello packetsat a visited vertex, while L = 1 is used for the zigzag patternwidth, unless otherwise noted.

A. Network Traversing

We investigate network exploration performance with re-spect to the number of UAVs. We measure network travers-ing time as complete coverage time by which all of UAVsterminate the traversing procedure, assuming that UAVs flywith the speed of 11.1 m/s (as per the specification of ParrotAR.Drone 2.0). As in Fig. 4(a), the complete coverage time

드론 수 1 2 4 6

avg #REF! #REF! 1.5407 #REF!

std #REF! 10.28786 7.497333 #REF!

duplicated #REF! 41.6 48.1 #REF! 2.806202

time 474.2481 245.8233 139.4682 105.2326

avg #REF! 3.0132 1.7608 #REF!

std #REF! 13.93413 12.04824 #REF!

duplicated #REF! 46.6 53.4 #REF!

time 554.5054 284.2682 159.3923 121.7892

avg 108.6939 70.48016 39.71151 28.00064

ant 고정

ant 랜덤

200

300

400

500

600

Net

work

Tra

vers

ing T

ime

(sec

)Ants (fixed)

Ants (random)

DroneNet (fixed)

DroneNet (random)

avg 108.6939 70.48016 39.71151 28.00064

std 0.826281 12.53909 7.124203 1.689763

duplicated 0 8 12.4 12.4

time 307.2791 201.4853 118.1411 83.90543

avg 3.359586 2.034884 1.174679 0.839662

std 2.736431 12.43376 6.71116 5.343019

duplicated 0 9.7 10.8 16.6

time 304.7535 187.1736 110.0031 83.90543

L=1 고정

L=1 랜덤0

100

200

300

400

500

600

1 2 4 6

Net

work

Tra

vers

ing T

ime

(sec

)

# of UAVs

Ants (fixed)

Ants (random)

DroneNet (fixed)

DroneNet (random)

(a) Complete coverage time per UAV for network traversing.

드론 수 1 2 4 6

avg #REF! #REF! 1.5407 1.24

std #REF! 10.28786 7.497333 0

duplicated #REF! 41.6 48.1 #REF!

time 169 87.6 49.7 37.5

avg #REF! 3.0132 1.7608 #REF!

std #REF! 13.93413 12.04824 #REF!

duplicated #REF! 46.6 53.4 #REF!

time 197.6 101.3 56.8 43.4

avg 108.6939 70.48016 39.71151 28.00064

ant 고정

ant 랜덤

30

40

50

60

70

80

90

100

3

4

5

6

7

Dup

licat

e Cove

rage

(%)

Trav

el D

ista

nce

(km

)

Ants (TD)DroneNet (TD)Ants (DC)DroneNet (DC)

avg 108.6939 70.48016 39.71151 28.00064

std 0.826281 12.53909 7.124203 1.689763

duplicated 0 8 12.4 12.4

time 109.5 71.8 42.1 29.9

avg 3.359586 2.034884 1.174679 0.839662

std 2.736431 12.43376 6.71116 5.343019

duplicated 0 9.7 10.8 16.6

time 108.6 66.7 39.2 29.9

40

L=1 고정

L=1 랜덤 0

10

20

30

40

50

60

70

80

90

100

0

1

2

3

4

5

6

7

1 2 4 6

Dup

licat

e Cove

rage

(%)

Trav

el D

ista

nce

(km

)

# of UAVs

Ants (TD)DroneNet (TD)Ants (DC)DroneNet (DC)

(b) Travel distance and duplicate coverage rate per UAV.

Fig. 4. Network exploration performance compared to Ants algorithm, withrespect to the number of UAVs.

of a UAV decreases as the number of UAVs increases forboth algorithms, while our DroneNet outperforms Ants with afactor of up to 1.8. We also explore how the initial locationof each UAV affects the performance. We let each UAVinitially be located at a fixed vertex or a randomly selectedvertex and launch its exploration. Our simulation result showsaverage performance over 10 runs. DroneNet shows verystable performance irrespective of the initial location of UAVs,whereas Ants is relatively more sensitive to where the UAVsinitiate their exploration. We also measure the average traveldistance of a UAV as in Fig. 4(b), showing a similar patternas network traversing time.

To understand how DroneNet can result in small traversingtime and distance, we measure how many vertexes each UAVhas redundantly explored over the total number of visitedvertexes as the duplicate coverage rate. As shown in Fig. 4(b),DroneNet keeps a small duplicate coverage rate around 10%with two and four UAVs, achieving less than 20% even with 6UAVs, whereas Ants leads to almost 50% redundant coverageperformance. This means that our motion planning algorithmminimizes the number of redundant vertex visits even if it isfully distributed, while guaranteeing complete coverage.

B. Network Recovery Performance

We investigate network recovery performance in terms ofend-to-end routing cost with respect to the number of UAVs

0 5 10 15 20 25 30 35 40 450

10

20

30

40

50

60

70

80

90

100

Routing Cost

Cum

ulat

ive

Fra

ctio

n (%

)

Before DeploymentDroneNet (1 UAV)DroneNet (2 UAVs)DroneNet (4 UAVs)DroneNet (6 UAVs)

(a) Cumulative distribution of routing cost for before vs. afterUAV deployment.

0 1 2 4 60

10

20

30

40

50

60

# of Deployed UAVs

Per

cent

age

of s

rc−

dst P

airs

with

No

Rou

te (

%)

(b) Routing hole percentage with respect to the number ofdeployed UAVs.

Fig. 5. Network recovery performance in terms of end-to-end routing costwith respect to the number of deployed UAVs

0

5

10

15

20

25

Per

cent

age

of s

rc−

dst P

airs

with

No

Rou

te (

%)

w/o Opt with 4 UAVs w/ Opt with 4 UAVs(DroneNet)

Fig. 6. Routing hole percentage for a heuristic algorithm without optimizationvs. our DroneNet optimization algorithm.

to deploy. The end-to-end routing cost is defined as thesummation of per-hop link costs where the per-hop link costis calculated as the expected number of transmissions on thelink.

Fig. 5 shows that as a larger number of UAVs are deployed,the source-to-destination routing cost decreases. Given thesimulated network topology, the most effective number ofUAVs turns out to be 4. In other words, two more UAVsbeyond 4 UAVs do not greatly improve routing performance

0

0.2

0.4

0.6

0.8

1

1.2

0

5

10

15

20

25

30

35

40

45

50

7x7 10x10

Rela

tive

Com

munic

ation O

verh

ead

Perc

enta

ge o

f sr

c-dst

Pairs

with N

o R

oute

(%

)

Total # of Vertexes over ROI, M (=mxm)

Percentage of src-dst Pairswith No Route (%)

Relative CommunicationOverhead

Fig. 7. Impact of sampling grid size on network communication overhead

further. This implies that depending on a sparsity of networktopology, there exists an effective number of UAVs for themost influential routing enhancement.

We compare our deployment optimization algorithm againsta heuristic deployment algorithm without optimization. Forthis comparison, both algorithms use the same PRR tableachieved after network traversing. The only difference is thatthe heuristic algorithm deploys UAVs into the vertexes withthe lowest PRRs excluding zero, whereas our algorithm appliesa binary integer program-based optimization that reducesoverlapped radio range among the deployed UAVs. As Fig. 6demonstrates, our optimization technique lessens the numberof source-to-destination pairs with no route (i.e., routing holes)with a factor of 1.6, compared to a heuristic approach.

Lastly, we would like to discuss how the probing densityaffects routing performance and communication overhead. Wemeasure communication overhead as the accumulated packettransmissions for sending hello packets and response packetsfrom each UAV, and exchanging the vertex-visit-list and thePRR table among UAVs. As shown in Fig. 7, as the probingdensity increases from 7 × 7 to 10 × 10, approximately by2, we can achieve routing performance improvement with afactor of 3.1, while consuming more communication overheadwith a factor of 2.3. This demonstrates that DroneNet canachieve a higher benefit of network-wide data delivery witha relatively smaller network overhead increase, showing aninteresting trade-off relationship.

VI. CONCLUSION

We have presented a self-organizing UAV deployment al-gorithm based on sparse network status probing from the airalong with a distributed motion planning. We have designeda network traversing algorithm based on a fully distributedlocal decision for its next movement, while minimizing theduplicate coverage and guaranteeing complete coverage. Then,we have formulated the problem of UAV relay deployment intoa binary integer program that aims to spread them into thearea under somewhat wide link outage, reducing overlappedcoverage makeup.

Our experiment results indicate that our network traversingalgorithm covers the complete ROI more effectively in terms

of coverage time, travel distance, and duplicate coverageagainst a popularly used multi-agent exploration algorithmAnts. Also, our optimization technique finds an efficient de-ployment to repair the routing hole problem, while slightlyincreasing network overhead in return.

For future work, we would devise a UAV deployment algo-rithm that fundamentally infers the overall network topologybased on only a few connectivity probing. By applying thecompressive sensing signal processing theory with networktomography, we may obtain a more globally optimal solutionfor a drastic network reconstruction with only few UAVs. Also,the optimal motion planning of UAVs considering rechargingin case of out-of-battery scenarios would be an interestingresearch direction.

ACKNOWLEDGMENT

This work was supported by the National Research Foun-dation of Korea funded by the Korean Ministry of Sci-ence, ICT, and Future Planning through the Basic ScienceResearch Program under Grants NRF-2013R1A1A1009854,NRF-2013R1A2A2A04014772.

REFERENCES

[1] N. Ahmed, S. S. Kanhere, and S. Jha. The holes problem in wirelesssensor networks: a survey. ACM SIGMOBILE Mobile Computing andCommunications Review, 9(2):4–18, 2005.

[2] M. Asadpour, D. Giustiniano, and K. A. Hummel. From ground to aerialcommunication: dissecting wlan 802.11 n for the drones. In Proceedingsof the 8th ACM international workshop on Wireless network testbeds,experimental evaluation & characterization, pages 25–32. ACM, 2013.

[3] J. L. Bredin, E. D. Demaine, M. Hajiaghayi, and D. Rus. Deployingsensor networks with guaranteed capacity and fault tolerance. InProceedings of the 6th ACM international symposium on Mobile adhoc networking and computing, pages 309–319. ACM, 2005.

[4] P. Corke, S. Hrabar, R. Peterson, D. Rus, S. Saripalli, and G. Sukhatme.Autonomous deployment and repair of a sensor network using an un-manned aerial vehicle. In Robotics and Automation, 2004. Proceedings.ICRA’04. 2004 IEEE International Conference on, volume 4, pages3602–3608. IEEE, 2004.

[5] M. Di Felice, A. Trotta, L. Bedogni, K. R. Chowdhury, and L. Bononi.Self-organizing aerial mesh networks for emergency communication. InProceedings of the 25th IEEE International Symposium on Personal,Indoor, and Mobile Radio Communications (PIMRC 2014 - Mobile andWireless Networks), 2014.

[6] E. Ferranti, N. Trigoni, and M. Levene. Brick& mortar: an on-line multi-agent exploration algorithm. In Robotics and Automation, 2007 IEEEInternational Conference on, pages 761–767. IEEE, 2007.

[7] E. Ferranti, N. Trigoni, and M. Levene. Rapid exploration of unknownareas through dynamic deployment of mobile and stationary sensornodes. Autonomous Agents and Multi-Agent Systems, 19(2):210–243,2009.

[8] A. Goldsmith. Wireless Communications. Cambridge University Press,New York, NY, USA, 2005.

[9] N. Hazon, F. Mieli, and G. A. Kaminka. Towards robust on-linemulti-robot coverage. In Robotics and Automation, 2006. ICRA 2006.Proceedings 2006 IEEE International Conference on, pages 1710–1715.IEEE, 2006.

[10] V. Isler, S. Kannan, and K. Daniilidis. Sampling based sensor-networkdeployment. In Intelligent Robots and Systems, 2004.(IROS 2004).Proceedings. 2004 IEEE/RSJ International Conference on, volume 2,pages 1780–1785. IEEE, 2004.

[11] J. Svennebring and S. Koenig. Building terrain-covering ant robots: Afeasibility study. Autonomous Robots, 16(3):313–332, 2004.

[12] E. Yanmaz, R. Kuschnig, and C. Bettstetter. Achieving air-ground com-munications in 802.11 networks with three-dimensional aerial mobility.In INFOCOM, 2013 Proceedings IEEE, pages 120–124. IEEE, 2013.